DeepSeek and Moonshot were quietly relaying customer prompts to Claude (twitter.com)

In a startling revelation, Chinese labs DeepSeek and Moonshot have been found to be secretly forwarding customer prompts to Anthropic’s AI model, Claude, using fraudulent accounts. This deceptive practice misled users into believing they were interacting with local Chinese AI, when in reality, their inquiries—and sensitive data—were being relayed to Claude. Notably, one user with ties to the Chinese military sought AI insights on a person’s suspicious behavior tracked via extensive CCTV surveillance in Chengdu, while DeepSeek even transmitted live credentials from a Russian government database. This incident raises significant concerns within the AI/ML community regarding data privacy and ethical usage. The unauthorized access to such sensitive information highlights vulnerabilities in AI deployment and the potential misuse of AI technologies to engage in surveillance or military activities. Moreover, the sheer volume of prompts—nearly 300,000 in just ten days—indicates a troubling trend in how AI systems can inadvertently be co-opted for more nefarious purposes. The implications of this breach extend beyond individual privacy to challenge the integrity of AI training practices, demanding a critical reevaluation of data handling and accountability in AI applications.